Transfer Learning Improves Supervised Image Segmentation Across Imaging Protocols

被引:185
作者
van Opbroek, Annegreet [1 ,2 ]
Ikram, M. Arfan [3 ,4 ]
Vernooij, Meike W. [3 ,4 ]
de Bruijne, Marleen [1 ,2 ,5 ]
机构
[1] Univ Med Ctr Rotterdam, Erasmus MC, Biomed Imaging Grp Rotterdam, Dept Med Informat, NL-3000 CA Rotterdam, Netherlands
[2] Univ Med Ctr Rotterdam, Erasmus MC, Biomed Imaging Grp Rotterdam, Dept Radiol, NL-3000 CA Rotterdam, Netherlands
[3] Univ Med Ctr Rotterdam, Erasmus MC, Dept Radiol, NL-3000 CA Rotterdam, Netherlands
[4] Univ Med Ctr Rotterdam, Erasmus MC, Dept Epidemiol, NL-3000 CA Rotterdam, Netherlands
[5] Univ Copenhagen, Dept Comp Sci, DK-2100 Copenhagen, Denmark
关键词
Image Segmentation; machine learning; magnetic resonance imaging; pattern recognition; transfer learning; WHITE-MATTER LESIONS; TISSUE CLASSIFICATION; BRAIN-TISSUE; INTENSITY NONUNIFORMITY; AUTOMATIC SEGMENTATION; MR-IMAGES; FRAMEWORK; RISK;
D O I
10.1109/TMI.2014.2366792
中图分类号
TP39 [计算机的应用];
学科分类号
081203 ; 0835 ;
摘要
The variation between images obtained with different scanners or different imaging protocols presents a major challenge in automatic segmentation of biomedical images. This variation especially hampers the application of otherwise successful supervised-learning techniques which, in order to perform well, often require a large amount of labeled training data that is exactly representative of the target data. We therefore propose to use transfer learning for image segmentation. Transfer-learning techniques can cope with differences in distributions between training and target data, and therefore may improve performance over supervised learning for segmentation across scanners and scan protocols. We present four transfer classifiers that can train a classification scheme with only a small amount of representative training data, in addition to a larger amount of other training data with slightly different characteristics. The performance of the four transfer classifiers was compared to that of standard supervised classification on two magnetic resonance imaging brain-segmentation tasks with multi-site data: white matter, gray matter, and cerebrospinal fluid segmentation; and white-matter-/MS-lesion segmentation. The experiments showed that when there is only a small amount of representative training data available, transfer learning can greatly outperform common supervised-learning approaches, minimizing classification errors by up to 60%.
引用
收藏
页码:1018 / 1030
页数:13
相关论文
共 43 条
[1]   Probabilistic segmentation of brain tissue in MR imaging [J].
Anbeek, P ;
Vincken, KL ;
van Bochove, GS ;
van Osch, MJP ;
van der Grond, J .
NEUROIMAGE, 2005, 27 (04) :795-804
[2]   Automatic segmentation of different-sized white matter lesions by voxel probability estimation [J].
Anbeek, P ;
Vincken, KL ;
van Osch, MJP ;
Bisschops, RHC ;
van der Grond, J .
MEDICAL IMAGE ANALYSIS, 2004, 8 (03) :205-215
[3]  
[Anonymous], 2004, P 21 INT C MACH LEAR
[4]  
[Anonymous], 2001, J. Am. Stat. Assoc.
[5]  
[Anonymous], 2008, Midas J
[6]  
[Anonymous], 2007, Proc. ACM Int. Conf. on Multimedia
[7]   Unified segmentation [J].
Ashburner, J ;
Friston, KJ .
NEUROIMAGE, 2005, 26 (03) :839-851
[8]   LIBSVM: A Library for Support Vector Machines [J].
Chang, Chih-Chung ;
Lin, Chih-Jen .
ACM TRANSACTIONS ON INTELLIGENT SYSTEMS AND TECHNOLOGY, 2011, 2 (03)
[9]   A fully automatic and robust brain MRI tissue classification method [J].
Cocosco, CA ;
Zijdenbos, AP ;
Evans, AC .
MEDICAL IMAGE ANALYSIS, 2003, 7 (04) :513-527
[10]  
CORTES C, 1995, MACH LEARN, V20, P273, DOI 10.1023/A:1022627411411